search

MengzhangLI / STFGNN

Code of STFGNN@AAAI-2021 (Spatial-Temporal/ Traffic data forecasting)
210 stars 56 forks source link

About Missing Data (0 values) #15

Closed imethanlee closed 3 years ago

imethanlee commented 3 years ago

There is a lot of missing data (those with value 0) in several datasets (such as METR-LA). Did you adopt any interplolation approach (such as linear interplolation) to fill in these missing values? an example of missing values in METR-LA

MengzhangLI commented 3 years ago

There is a lot of missing data (those with value 0) in several datasets (such as METR-LA). Did you adopt any interplolation approach (such as linear interplolation) to fill in these missing values? an example of missing values in METR-LA

Hi, sorry for late reply.

First of all, the result of this model on METR-LA is wrong because the metric of METR-LA, PEMS-BAY, etc is different from PEMS0X. Sorry for my misleading result on first version on ARXIV.

But I can answer this question: right now most model use masked index to solve many zero/null values in METR-LA and PEMS-BAY. When training, the masked index could skip places of those zeor/null to get loss for back propagation.

For example, you can check it from GraphWaveNet: https://github.com/nnzhan/Graph-WaveNet/blob/6b162e80c59a1d494809252eca055cff93dc66b1/util.py#L177

Best,

imethanlee commented 3 years ago

This helps a lot. Thanks.